Tiangong Auto Research for WorkBuddy

For feedback or Issue requests, load the sibling tiangong-auto-research/SKILL.md and follow its reporting route before workspace inspection. Its installed references/issue-reporting.md supplies the common templates and reporting rules.

Use the research workflow below only inside a user-selected managed Auto Research workspace. Inspect the exact workspace context first. If setup is absent, use the canonical Skill's setup reference; do not copy control files or invent a workspace.

Load the sibling project-installed tiangong-auto-research/SKILL.md, then read its references/sandboxed-ide.md and the other reference selected by the current stage. That canonical Skill owns the research workflow, schemas, budgets, evidence rules, and recovery behavior; this adapter does not duplicate them.

Run canonical CLI operations through this adapter's scripts/workbuddy_research_cli.sh. It selects only an explicit Node.js 24 runtime and then delegates to the sibling canonical resolver. Do not call WorkBuddy's ambient Node 22/npx after an EBADENGINE warning and do not suppress the engine requirement.

Keep WorkBuddy in Default Permission mode. Never choose Full Access, an unsandboxed-command escape hatch, dangerouslyDisableSandbox, dangerously-skip-permissions, excludedCommands, or silent transport fallback.

The workspace must identify the actual native producer as workbuddy or codebuddy. Use that same value for the packet's --host-agent; never impersonate Codex or Claude. The CLI must not launch the producer as a child.

For sandbox-bridge, require the owner to start the exact-version reviewer sidecar from a separate native terminal and a private state directory outside the workspace. From the IDE, verify research reviewer status, then run the explicitly cost-confirmed reviewer doctor. Stop on any structured bridge, signature, nonce, model, version, result-binding, or sandbox-policy failure.

After readiness passes, perform native producer stages in this current task and return every result through the canonical CLI packet. Independent review and mechanical closure remain CLI-controlled.